This is research code for the approximate GP model FactorGP, submitted to the RSS From Perception to Action workshop. Our model learns a GP approximating GMRF prior that can encode spatially varying lengthscales and conditional independences, capturing non-stationarity and discontinuities in data respectively.
The model code is contained within fgp/:
fgp/graph_learning.py: SVGD-based graph learning modulefgp/gbp.py: factor graph representation of the GMRF prior
- Python ≥ 3.9
- Other dependencies:
pip install -r requirements.txt
python active_mapping_srtm.py --env <environment_name> [--samples N]
Valid environments: N17E073, N43W080, N45W123, N47W124
Example:
python active_mapping_srtm.py --env N17E073
python active_mapping_srtm.py --env N43W080 --samples 700
Results are saved as results_<env>_seed<N>.png in the working directory and the eval table is printed to terminal.
python active_sampling_gas.py
Per-step figures and final metrics are saved under figs/active_sampling/ and timestamped run data under results/active_sampling/. Eval table is printed to terminal.
Code to produce gas fields via a forward-Euler Fick's law solver is in data/gas/diffusion_test.py. You can define your own building floor plan and source/sink locations to produce new simulated gas fields.